PaperlayerView exact revision on arXiv

Source-linked AI summary

Hybrid Recommender Systems: A Systematic Literature Review

Erion Çano, Maurizio Morisio

arXiv:1901.03888v1cs.IRcs.CYcs.LG

TL;DR

Hybrid recommender research lacked a focused quantitative synthesis of its challenges, techniques, applications, and evaluation practices. This paper conducts a systematic literature review of recent hybrid recommenders, finding recurring use of collaborative filtering and persistent evaluation limitations alongside established and emerging research directions.

  • Problem

    The literature lacked a quantitative systematic review focused entirely on recent hybrid recommender systems and their problems, techniques, applications, and evaluations.

  • Method

    The paper systematically analyzes 76 primary studies using a reproducible review methodology and synthesizes their problems, techniques, hybridization classes, domains, datasets, evaluations, and future work.

  • Results

    Most studies combine collaborative filtering with another technique, while cold-start, data sparsity, and accuracy are the most recurrent problems.

  • Takeaways & Limitations

    Hybrid recommenders remain a broad basis for addressing established and newer challenges such as context variation, evolving tastes, and cross-domain recommendation.

  • Takeaways & Limitations

    Comparisons with similar methods and accuracy metrics often provide limited or non-credible evidence about recommendation or system quality.

Abstract

from arXiv · show

Recommender systems are software tools used to generate and provide suggestions for items and other entities to the users by exploiting various strategies. Hybrid recommender systems combine two or more recommendation strategies in different ways to benefit from their complementary advantages. This systematic literature review presents the state of the art in hybrid recommender systems of the last decade. It is the first quantitative review work completely focused in hybrid recommenders. We address the most relevant problems considered and present the associated data mining and recommendation techniques used to overcome them. We also explore the hybridization classes each hybrid recommender belongs to, the application domains, the evaluation process and proposed future research directions. Based on our findings, most of the studies combine collaborative filtering with another technique often in a weighted way. Also cold-start and data sparsity are the two traditional and top problems being addressed in 23 and 22 studies each, while movies and movie datasets are still widely used by most of the authors. As most of the studies are evaluated by comparisons with similar methods using accuracy metrics, providing more credible and user oriented evaluations remains a typical challenge. Besides this, newer challenges were also identified such as responding to the variation of user context, evolving user tastes or providing cross-domain recommendations. Being a hot topic, hybrid recommenders represent a good basis with which to respond accordingly by exploring newer opportunities such as contextualizing recommendations, involving parallel hybrid algorithms, processing larger datasets, etc.

1. Introduction

Recommender systems emerged to help users navigate information overload, while hybrid systems combine complementary recommendation strategies to offset individual limitations. This review systematically synthesizes hybrid recommender research from the previous decade.

  • Information overload from expanding digital information created demand for tools that help users choose among many alternatives.
  • Recommender systems provide suggestions for items and other recommendable entities using strategies such as collaborative, content-based, demographic, and knowledge-based filtering.
  • Collaborative filtering uses user feedback and similarities, whereas content-based filtering uses item features and user profiles.
  • Hybrid systems combine two or more strategies to reinforce advantages and reduce disadvantages, exemplified by Fab’s combination of collaborative and content-based filtering.
  • The review applies a systematic methodology to synthesize hybrid recommender research and examines problems, techniques, applications, evaluation, and future directions.

2. Methodology

The review used a rigorous, reproducible protocol to identify, select, assess, extract, and synthesize evidence about hybrid recommender systems. Its research questions covered problems, techniques, hybridization, domains, evaluation, datasets, and future work.

  • The review followed Kitchenham and Charters’ systematic-literature-review guidelines to support verifiable and unbiased research synthesis.
  • The researchers defined questions covering relevant studies, problems, mining techniques, recommendation combinations, hybridization classes, domains, evaluation, datasets, and future directions.
  • The search used five digital libraries and an expanded keyword string built from hybrid, recommender, and system-related terms.
  • 9673 preliminary studies were retrieved, then inclusion and exclusion criteria reduced the corpus to 76 journal and conference papers.
  • Study quality was assessed with six questions scored 0, 0.5, or 1 and weighted by importance coefficients of 0.5, 1, or 1.5.
  • Data extraction collected bibliographic and research-content information, which was organized and synthesized thematically using codes and sub-codes in Nvivo.

3. Results

The results section reports findings from the selected studies across problems, techniques, hybridization classes, evaluation methodologies, and related categories. It uses examples from included studies and discusses the results afterward.

  • The section presents findings from the selected studies to answer each research question.
  • Results are organized into categories including problems, techniques, hybridization classes, and evaluation methodologies.
  • Examples from included studies illustrate the reported categories, with broader discussion provided in the following section.

3.1. RQ1: Included studies

The review included 76 hybrid recommender papers published from 2005 to 2015, with most appearing in the final five years. Journal studies generally received higher quality scores than conference studies.

  • 76 hybrid recommender papers published from 2005 to 2015 were selected for further processing.
  • About 75% of journal studies had quality scores above 0.9, compared with about 35% of conference studies.
  • QQ4 had the highest average score at 0.947, while QQ3 had the lowest at 0.651.
  • Every included paper had a weighted quality score above 0.81, and one journal study achieved the maximum score of 1.0.

3.2. RQ2: Research problems

The review identifies 12 problems addressed by hybrid recommender studies, with cold-start and data sparsity among the most frequent. Accuracy, scalability, diversity, and less frequent concerns such as privacy and adaptiveness also receive attention.

  • Cold-start: 23 studies address cold-start recommendations for new users or items.Proposed remedies include latent-feature pseudo-ratings, CF-CBF combinations, and weighted strategy aggregation.
  • Data sparsity: Data sparsity arises from limited ratings, producing insufficient information for identifying similar users or items.It is especially prevalent in collaborative filtering and is addressed with cross-domain factorization models.
  • Accuracy: 16 studies target recommendation accuracy, including hybrid CF-CBF models that report improved recommendation quality.Accuracy and data sparsity appear together in six studies.
  • Scalability: Scalability was addressed in 11 studies through methods such as dataset reduction and hybrid CF-CBF filtering.The Hyred system reduced the dataset, alleviated sparsity, and slightly reduced computational time.
  • Diversity: Diversity is pursued to reduce popularity bias and avoid recommendation lists containing highly similar items.K-Furthest Neighbors was used to create more diverse recommendations.
  • Other problems: Other reported problems include lack of personalization, privacy preserving, noise reduction, data source integration, lack of novelty, and user preference adaptiveness.These concerns appear in relatively few studies.

3.3. RQ3a: Data mining and machine learning techniques

The reviewed hybrid recommenders use a wide variety of data mining and machine learning techniques, often assigning different techniques to different solution components. K-NN is the most frequent identified technique, while matrix manipulation, clustering, association rules, fuzzy logic, and other methods also appear.

  • K-NN: K-NN appears in 59 studies and is mainly used to analyze neighborhoods of similar users or items.It is particularly common among collaborative filtering recommenders.
  • Clustering: Clustering is typically applied to preprocess data, with examples combining K-means variants and CBF to build e-learning recommenders.One reported application achieved improved accuracy for learning-item recommendations.
  • Association rules: Association-rule mining discovers item relations expressed as X => Y and can derive recommendations such as “X likes Y.”A discussion-group application used association rules to form user-similarity neighborhoods.
  • Fuzzy logic: Fuzzy logic provides item-aware methods that complement collaborative filtering and can be combined with CF systems.One example combined two CF systems with a fuzzy inference system in a weighted way to achieve better accuracy.
  • Matrix manipulation: Matrix manipulation techniques appear in 9 studies, including SVD, LDA, PCA, dimensionality reduction, and matrix factorization.Examples use LDA for rating probabilities and dimensionality reduction for sparsity and scalability.
  • Other techniques: Other techniques, including genetic algorithms, Naive Bayes, neural networks, expert notions, and statistical modeling, appear in 19 papers.These methods represent less frequent approaches in the review.

3.4. RQ3b: Recommendation technique combinations

Hybrid recommendation studies combine recommendation strategies in several patterns, most prominently collaborative filtering with content-based filtering or another technique. These combinations target problems including performance, scalability, data sparsity, and accuracy across varied domains.

  • CF with other techniques: CF combined with techniques other than CBF is used to improve performance or incorporate demographic information.Examples include clustering with CF for multi-criteria hotel recommendation and fuzzy demographic similarities integrated with user-based similarities.
  • CF-CBF: CF-CBF is a highly popular combination, often producing predictions through weighted outputs from both strategies.Some systems instead switch between CF and CBF or use more complex combinations.
  • CF-CBF: A CF-CBF system reduced the dataset with Pearson-based neighborhoods, then combined CF and distance-to-boundary CBF scores linearly.The authors reported improved accuracy compared with algorithms using full datasets.
  • CF-CBF: Bayesian-network CF-CBF hybrids automatically select strategy weights according to problem conditions and report improved recommendation accuracy.The model computes probability distributions over expected ratings.
  • Three-way hybrids: Some studies combine CF and CBF with a third approach, such as demographic filtering, to recommend for groups of similar profiles in data-sparse settings.The described staged approach analyzes ratings and item attributes before invoking CF.
  • Other combinations: Other combinations include Item-Item CF with User-User CF, CBF with non-CF techniques, and hybrids using trust, association rules, neural networks, or semantic methods.The review also reports combinations involving genetic algorithms, dimensionality reduction, social tagging, ontologies, pattern mining, and classifiers.

3.5. RQ4: Classes of hybridization

The review classifies hybrid recommender systems by how recommendation techniques are combined, with weighted hybrids occurring most frequently. Other classes differ in sequencing, data flow, or aggregation, and several offer specific trade-offs in performance, diversity, simplicity, or implementation effort.

  • 3.5.1. Weighted: Weighted hybrids were the most frequent class, aggregating recommendation-technique scores with weighted linear functions.P-Tango combined collaborative-filtering and content-based scores, with per-user weights adapted through user feedback.
  • 3.5.3. Cascade: Cascade hybrids use a staged process in which one technique ranks candidate items and a second refines the preliminary list.Because cascades are order-sensitive, reversing the technique order can produce different recommendations.
  • 3.5.4. Switching: Switching hybrids select among recommendation techniques according to criteria such as recommendation credibility or classifier confidence.One example switches from Naive Bayes recommendations to item-item collaborative filtering when confidence is poor.
  • 3.5.5. Feature augmentation: Feature augmentation hybrids feed one recommender’s prediction or classification into another as additional item or feature data.This class appeared in 7 (9.2%) studies, and one method increased diversity with only minimal accuracy loss.
  • 3.5.6. Meta level: Meta-level hybrids use an entire model produced by one technique as input to a second technique.The class was found in 7 (9.2%) studies; its integration effort is considerable and often requires advanced constructs.
  • 3.5.7. Mixed: Mixed hybrids combine the item lists generated by multiple recommenders into a final list and were found in 3 (3.9%) studies.Their simplicity can help address cold-start problems, while overlapping recommendations may be highlighted as especially interesting.

3.6. RQ5: Application domains

The review identified 18 application domains for hybrid recommenders, alongside domain-independent studies. Movies were the most common specific domain, followed by education or e-learning, music, and web services.

  • Application-domain distribution: 18 application domains were identified, while 21 (27.6%) studies were domain independent.Domain-independent methods were not limited to a particular domain and could be applied with minor or no changes.
  • Application-domain distribution: 17 (22.3%) studies applied hybrid recommenders to the movie domain.Public movie-preference feedback and movie datasets were identified as helpful resources for this concentration.
  • Application-domain distribution: 9 (11.8%) studies addressed education or e-learning, and 6 (7.8%) addressed music.These were the next most frequent named application domains after movies.

3.7. RQ6: Evaluation

The review examines how hybrid recommender systems are evaluated, which recommendation characteristics and metrics are measured, and which datasets support experiments. Evaluations predominantly compare similar methods using accuracy measures, while user surveys and broader datasets are less common.

  • Evaluation methodologies: 58 studies compared proposed systems with similar methods, usually contrasting hybrid systems with pure CF or CBF.Accuracy and error metrics such as MAE and RMSE were common in these comparisons.
  • Characteristics and metrics: 62 studies evaluated accuracy, followed by user satisfaction in 10 studies, diversity in 7, computational complexity in 6, and novelty or serendipity in 4.Precision, recall, and F1 were the most frequent accuracy metrics, while surveys assessed user satisfaction.
  • Datasets: 55 studies used at least one public dataset, whereas 21 used no dataset.Some studies without datasets relied on synthetic data, surveys, or other techniques.
  • Datasets: MovieLens was used in 26 studies, making it the most popular named public dataset in the review.Other movie datasets included EachMovie in 6 studies, FilmTrust in 5, and Yahoo-Movie in 3.

3.8. RQ7: Future work

The review identifies future work centered on extending existing systems, improving evaluation, incorporating context and data, and exploring alternative algorithms, hybridization classes, domains, and objectives.

  • Future work directions: Extending proposed solutions was the most common future-work direction, appearing in 14 studies (18.4%).Suggested extensions included additional algorithms, greater personalization, more datasets, and more item features.
  • Future work directions: 8 studies (10.5%) suggested adding contextual data such as location and time of day to recommendations.
  • Future work directions: 7 studies (9.2%) proposed using more data or item features, while 6 studies (7.9%) proposed different algorithms or similarity measures.
  • Future work directions: Five studies (6.5%) proposed trying another hybridization class, and 7 studies (9.2%) suggested considering other application domains.

4. Discussion

The discussion finds that hybrid recommender research is growing and commonly combines collaborative filtering with other techniques to address recurrent problems. Evaluation remains dominated by offline accuracy comparisons, while future work emphasizes broader assessment and contextual, scalable, and cross-domain systems.

  • Research trends: More than 76% of included papers were published in the second half of the reviewed decade, indicating sustained recent research activity.
  • Problems and challenges: Cold-start, data sparsity, and accuracy are the most recurrent problems addressed by hybrid approaches.Hybrid systems combine techniques that exploit item features, ratings, or other information to address these problems.
  • Techniques and strategies: K-NN is the most popular data-mining technique, while CF is the most commonly used recommendation strategy.Clustering and association rules are also frequently used in hybrid systems.
  • Hybridization classes: Weighted hybridization is the most popular class, whereas mixed hybrid is least common and 13.2% of studies fall into an unclassified “Other” category.Weighted hybrids often combine CF and CBF dynamically, while some classes require more complex aggregation.
  • Application domains and datasets: Movies are the most common recommendation domain, partly because public movie-preference data are widely available.Movie datasets were used in more than 72% of the studies.
  • Evaluation: Most evaluations compare similar methods offline using accuracy or error metrics, while user surveys are less common but more direct and credible.The review characterizes comparison-based evaluation as often hermetic and not credible enough.
  • Future directions: Future directions include context-aware recommendations, better evaluation, more data, parallel hybrid algorithms, and cross-domain recommenders.

5. Conclusions

This review analyzes 76 primary studies to characterize hybrid recommender problems, techniques, hybridization classes, domains, datasets, evaluation, and future directions. It finds collaborative filtering and weighted hybrids dominant, while evaluation credibility and contextual, scalable, cross-domain recommendation remain open directions.

  • Scope and contribution: 76 primary studies were analyzed across problems, techniques, recommendation combinations, hybridization classes, domains, datasets, evaluation, and future work.
  • Main findings: Cold-start, data sparsity, and accuracy are the most recurrent problems addressed by hybrid approaches.Association rules, matrix factorization, and fuzzy logic are among the techniques used to address these problems.
  • Main findings: Collaborative filtering was the most commonly used recommendation strategy, and K-NN was the most popular data-mining technique.
  • Main findings: Weighted hybridization was the most recurrent class, while meta-level and feature-augmentation hybrids were rare.The review attributes the weighted class’s prevalence to its simplicity and dynamicity.
  • Evaluation: Accuracy dominated evaluation, which predominantly used comparisons with similar methods and error or prediction metrics.The review describes this methodology as often hermetic and not credible, while surveys are more user oriented.
  • Future directions: Future work emphasizes context awareness, rapid shifts in user interest, larger datasets, parallel algorithms, cross-domain recommendation, and lower computational complexity.

Appendix A. Selected Papers

Appendix A lists selected papers on hybrid recommender systems, covering publications from 2007 to 2014 across multiple recommendation settings and application domains.

  • Hybrid approaches: The appendix includes methods combining collaborative filtering with content, ontology, fuzzy, topic-model, and machine-learning techniques.Representative titles address content-based and collaborative recommendations, ontology-supported recommendation, fuzzy profiles, topic models, and switching hybrids using classifiers.
  • Research targets: Several selected papers explicitly target cold-start, personalized profiles, predictive accuracy, and recommendation efficiency.Examples include a method for the new-user cold-start problem, predicted-profile content recommendation, and optimization of combined collaborative recommender systems.
  • Publication coverage: The selected papers include hybrid recommender approaches published between 2007 and 2014.Examples include works from 2007, 2010, 2011, 2012, 2013, and 2014.
  • Application domains: Application domains represented include music, e-learning, tourism, news, telecom products and services, images, web services, and Yelp preferences.The listed works cover both general recommender methods and domain-specific systems.
Loading 1901.03888v1…